Speech enhancement using STFT of real and imaginary parts of modulation signals

Abstract

This paper investigates an alternate modulation (RImodulation) AMS-based framework for speech enhancement, in which real and imaginary parts of the modulation signal are processed in secondary AMS procedures. We propose to apply MMSE magnitude estimation in this framework, and using subjective experiments, show that MMSE RI-modulation magnitude estimation produces stimuli which is preferred by listeners over RI-modulation spectral subtraction. Experiments presented also show that while this framework is suited to speech enhancement and offers theoretical advantages over the modulation AMS framework, resulting stimuli had similar quality to that produced by the corresponding modulation AMS-based method.